Calibration of microsimulation models using nonparametric statistical techniques

The calibration of traffic microsimulation models has received widespread attention in transportation modeling. A recent concern is whether these models can simulate traffic conditions realistically. The recent widespread deployment of intelligent transportation systems in North America has provided an opportunity to obtain traffic-related data. In some cases the distribution of the traffic data rather than simple measures of central tendency such as the mean, is available. This paper examines a method for calibrating traffic microsimulation models so that simulation results, such as travel time, represent observed distributions obtained from the field. The approach is based on developing a statistically based objective function for use in an automated calibration procedure. The Wilcoxon rank-sum test, the Moses test and the Kohmigorov-Smirnov test are used to test the hypothesis that the travel time distribution of the simulated and the observed travel times are statistically identical. The approach is tested on a signalized arterial roadway in Houston, Texas. It is shown that potentially many different parameter sets result in statistically valid simulation results. More important, it is shown that using simple metrics, such as the mean absolute error, may lead to erroneous calibration results.